Differences in aortic strain and diameter after endovascular and open thoracic aortic surgery

Objective The aim of this study was to evaluate changes in proximal aortic dynamics after thoracic aortic repair using a deep learning-based approach to automatically quantify aortic root diameter and strain from echocardiogram videos as a surrogate for aortic compliance. Methods This was a single-center retrospective cohort study with external cohort validation. We used 51,726 parasternal long-axis videos from Cedars-Sinai Medical Center to train an artificial intelligence segmentation model for frame-level measurement of the aortic root diameter and aortic strain. Model performance was evaluated on held-out internal data (n = 503) and externally validated on independent datasets from Stanford Healthcare (n = 1801) and the Unity Imaging Collaborative (n = 1440). We then used this tool in clinical cohorts to assess aortic strain in patients who underwent open and endovascular thoracic aortic repair as well as patients who had never undergone aortic surgical intervention. Surgical patients were adults treated for thoracic aortic disease, predominantly aortic aneurysm or dissection, with at least one technically adequate postoperative parasternal long-axis echocardiogram. Patients with prior aortic root reconstruction were excluded. Controls had no history of thoracic aortic surgical intervention on chart review and were drawn from the same institution. Surgical patients and controls were balanced using exact matching on age (in years, based on age at the first postoperative echocardiogram) and sex, 1:1 without replacement; surgical patients without an exact age- and sex-matched control were excluded from the matched analysis. The pooled surgical cohort comprised 794 patients, of whom 792 (99.7%) had an exact match and were carried into the matched analysis. Proximal landing zones for thoracic endovascular aortic repair (TEVAR) and the anatomic extent of open repairs are reported using the Ishimaru/Society for Vascular Surgery thoracic aortic zone classification (TEVAR proximal landing was most commonly at zones two or 3; open surgical repair [OSR] proximal extent was predominantly at zone 0); patients with isolated root replacement were excluded from the comparative analyses. Results The model achieved a mean absolute error (MAE) of 2.14 mm for aortic root diameter measurement (R 2 = 0.52) on internal test data, outperforming human inter-observer variability (MAE, 2.87 mm; P < .001). External validation showed consistent performance (Stanford MAE, 2.46 mm; Unity MAE, 1.88 mm). Aortic root diameters were significantly larger in both the TEVAR group and the OSR group compared with matched controls (both P < .001), with no significant difference between TEVAR and OSR ( P = .42). Aortic root strain was significantly higher in OSR than in matched controls ( P < .001) and was lower in TEVAR than in OSR ( P = .001), whereas TEVAR strain did not differ from controls ( P = .48). In secondary outcomes, patients with aortic regurgitation demonstrated elevated aortic strain, and no significant difference in strain was observed by proximal and distal zones of repair. Median time from surgery to the analyzed echocardiogram was 489.5 days (interquartile range, 8-1229 days) for TEVAR and 16 days (interquartile range, 5-345 days) for OSR. Conclusions A deep learning algorithm can measure aortic root diameter and identify changes in aortic strain with higher precision than human assessment from echocardiogram videos. The automated workflow allows for reproducible, precise measurements of aortic phenotypes to inform upstream aortic remodeling that may occur after open and endovascular aortic surgeries.